Crane Reproduction

This repository contains a reproducible Python implementation of the Crane scheduling workflow for DNN pipeline scheduling experiments.

What is included

• Layer and DAG modeling
• Hierarchical block construction and refinement
• ScT and MeT based scheduling / memory optimization
• Transformer and official NNS experiment entrypoints
• Standard text, CSV, and HTML outputs under outputs/experiments/

Environment

Python 3.11 is recommended.

Windows PowerShell

python -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install -r requirements.txt

Verify the environment

python --version
python -m pip --version

Main experiment entrypoints

Transformer hierarchical experiment

python example\run_transformer_min_layer_block_experiment.py

Transformer training reproduction

python example\run_transformer_training_repro.py --verbose-progress

Official NNS layer-level suite

python example\run_official_nns_layer_level.py

Output format

Each run writes a timestamped directory under outputs/experiments/, typically including:

• summary.txt
• summary.csv
• best_detail.txt
• best_detail.html
• network-specific detail files for batch runs

The standard outputs intentionally use repository-relative paths instead of machine-specific absolute paths.

Notes

• .venv, __pycache__, and generated experiment outputs are ignored by Git.
• The experiment scripts should be run from the repository root.
• Existing reference materials under outputs/docs/ are kept in the repository; generated experiment runs under outputs/experiments/ are not.